diff --git a/NEWS.md b/NEWS.md
index 80ce22f1..c597be0e 100644
--- a/NEWS.md
+++ b/NEWS.md
@@ -1,5 +1,6 @@
 # evaluate (development version)
 
+* evaluation "chunks" now provide a function-like scope. This means that `on.exit()` will now run at the end of the evaluate code, rather than immediately and `return()` will cause the evaluation to finish (#201).
 * The default `value` handler now evaluates print in a child environment of the evaluation environment. This largely makes evaluate easier to test, but should make defining S3 methods for print a little easier (#192).
 * `parse_all()` adds a `\n` to the end of every line, even the last one if it didn't have one in the input.
 * Setting `ACTIONS_STEP_DEBUG=1` (as in a failing GHA workflow) will automatically set `log_echo` and `log_warning` to `TRUE` (#175).
diff --git a/R/conditions.R b/R/conditions.R
index 37a7d97d..dee08f0e 100644
--- a/R/conditions.R
+++ b/R/conditions.R
@@ -51,8 +51,12 @@ with_handlers <- function(code, handlers) {
 }
 
 sanitize_call <- function(cnd) {
-  if (identical(cnd$call, quote(eval(expr, envir)))) {
+  if (identical(cnd$call, quote(withVisible(do)))) {
     cnd$call <- NULL
   }
+  if (identical(cnd$call, quote(eval(as.call(list(cb)), envir)))) {
+    cnd$call <- NULL
+  }
+  
   cnd
 }
diff --git a/R/evaluate.R b/R/evaluate.R
index da621617..63d4c955 100644
--- a/R/evaluate.R
+++ b/R/evaluate.R
@@ -134,34 +134,49 @@ evaluate <- function(input,
   # The user's condition handlers have priority over ours
   handlers <- c(user_handlers, evaluate_handlers)
   
-  for (tle in tles) {
-    watcher$push_source(tle$src, tle$exprs)
-    if (debug || log_echo) {
-      cat_line(tle$src, file = stderr())
-    }
+  cb <- function() {
+    do <- NULL # silence R CMD check note
 
-    continue <- withRestarts(
-      with_handlers(
-        {
-          for (expr in tle$exprs) {
-            ev <- withVisible(eval(expr, envir))
-            watcher$capture_plot_and_output()
-            watcher$print_value(ev$value, ev$visible, envir)
-          }
-          TRUE
-        },
-        handlers
-      ),
-      eval_continue = function() TRUE,
-      eval_stop = function() FALSE,
-      eval_error = function(cnd) stop(cnd)
-    )
-    watcher$check_devices()
-
-    if (!continue) {
-      break
+    for (tle in tles) {
+      watcher$push_source(tle$src, tle$exprs)
+      if (debug || log_echo) {
+        cat_line(tle$src, file = stderr())
+      }
+  
+      continue <- withRestarts(
+        with_handlers(
+          {
+            for (expr in tle$exprs) {
+              # Using `delayedAssign()` as an interface to the C-level function
+              # `Rf_eval()`. Unlike the R-level `eval()`, this doesn't create
+              # an unwinding scope.
+              eval(bquote(delayedAssign("do", .(expr), eval.env = envir)))
+              
+              ev <- withVisible(do)
+              watcher$capture_plot_and_output()
+              watcher$print_value(ev$value, ev$visible, envir)
+            }
+            TRUE
+          },
+          handlers
+        ),
+        eval_continue = function() TRUE,
+        eval_stop = function() FALSE,
+        eval_error = function(cnd) stop(cnd)
+      )
+      watcher$check_devices()
+  
+      if (!continue) {
+        break
+      }
     }
   }
+
+  # Here we use `eval()` to create an unwinding scope for `envir`.
+  # We call ourselves back immediately once the scope is created. 
+  eval(as.call(list(cb)), envir)
+  watcher$capture_output()
+
   # Always capture last plot, even if incomplete
   watcher$capture_plot(TRUE)
 
diff --git a/tests/testthat/test-conditions.R b/tests/testthat/test-conditions.R
index f760d835..abaf6d21 100644
--- a/tests/testthat/test-conditions.R
+++ b/tests/testthat/test-conditions.R
@@ -16,6 +16,10 @@ test_that("all condition handlers first capture output", {
 test_that("conditions get calls stripped", {
   expect_equal(evaluate("warning('x')")[[2]]$call, NULL)
   expect_equal(evaluate("stop('x')")[[2]]$call, NULL)
+
+  # including errors emitted by C
+  expect_equal(evaluate("mpg")[[2]]$call, NULL)
+  expect_equal(evaluate("3()")[[2]]$call, NULL)
 })
 
 test_that("envvar overrides keep_* arguments", {
diff --git a/tests/testthat/test-evaluate.R b/tests/testthat/test-evaluate.R
index 8a559e99..6bfbbe50 100644
--- a/tests/testthat/test-evaluate.R
+++ b/tests/testthat/test-evaluate.R
@@ -97,6 +97,24 @@ test_that("multiple lines of comments do not lose the terminating \\n", {
   expect_equal(ev[[1]]$src, "# foo\n")
 })
 
+test_that("on.exit is evaluated at end of code", {
+  ev <- evaluate::evaluate(c(
+    "on.exit(print('bye'))",
+    "print('hi')"
+  ))
+  expect_output_types(ev, c("source", "source", "text", "text"))
+})
+
+test_that("return causes an early return", {
+  ev <- evaluate::evaluate(c(
+    "1 + 1",
+    "return()", 
+    "2 + 2"
+  ))
+  expect_output_types(ev, c("source", "text", "source"))
+})
+
+
 test_that("check_stop_on_error converts integer to enum", {
   expect_equal(check_stop_on_error(0), "continue")
   expect_equal(check_stop_on_error(1), "stop")